a) the relation is slow slow because any associated geometries need to be built from the constituent relation members (just as way processing was slower than node processing).
b) the processing speeds up as the older relations tend to have more large ones and the fast to process ones (turn restrictions etc) tend to have become popular later on (you should have seen a similar effect during way processing).
c) I wouldn't set number-processors to 8 if you have 8 cores, you will have an additional database process per osm2pgsql thread and you may be overtaking overwhelming available memory
d) check that you are not swapping/paging (one time out is ok, back in not) and that you are not limited by the SAN, these days fast imports tend to run directly to local SSDs.